On Thu, Jul 16, 2009 at 11:57 AM, Jonathan
Schleifer<[email protected]> wrote:
> Am 16.07.2009 um 12:04 schrieb Kevin Smith:
>
>> If it takes 5 minutes for you to get initial presence from someone,
>> that means it's taking 5 minutes to establish s2s and do the presence
>> round-trip. Now, even if the server remembered (and told you) that
>> they're online, wouldn't that mean that the first message you sent
>> would also take the same time for the servers to get it in gear
>> together? In this case it might even be preferable see them as
>> offline, if you wouldn't be effectively able to communicate yet.
>
> No, the idea is to keep the s2s link open and the other server always sends
> presences, even if I'm not online. When I come online, my server could then
> send me the last presence.
>

I know a server (or two in fact) that don't aggressively close s2s
connections ;)

> The problem here is, as you already said, s2s. When I sign in, about 50 s2s
> links needs to be opened, which can take quite a while as my TLS key is 4096
> bits.
>
> If for example an s2s link would only be closed if it has been inactive for
> 48h, then you know it's really not going to be used anytime soon. I think
> keeping it for 48h and exchanging and caching presences would be ok and
> solve the issue. In fact, it would need less resources than the ultra-low
> timeout of 10 minutes that a lot of servers have.
>

I agree.

>> As an aside: I have no idea what would cause a server to consistently
>> take 5 minutes to get presence to you (apart from sever. overloading,
>> or brokenness) - does anyone have any thoughts on this, because it
>> seems to me that this is the underlying problem we're trying to solve
>> here.
>
> It's because all s2s links are opened at once, I guess.
>

I don't see why opening 50 connections should take 5 minutes. I
haven't tested the existing one, but with the new connection framework
we are writing for Prosody it can accept about 10,000 connections in
less than a few seconds. There is no reason a connection ought to take
5 minutes to complete, unless your s2s is using RFC 1149/2549.

> IMO, we need to specify a lot more for s2s - a lot there is unspecified. For
> example after how much time a connection should be considered idle.

I think this should be decided by the server admin, protocols
shouldn't dictate these things.

The issue of caching presences... it is possible to do that, I just
don't think it is necessary (read Kev's initial post to learn why).
The s2s connection issues are for the affected server implementations
to solve, not the XSF.

Matthew

Reply via email to